Off-line cursive handwriting recognition
In June 2005 I started as a PhD. student on the problem of off-line cursive handwriting recognition. Now some people asked: 'Why do you change from robots to handwriting recognition?'. And my answer is that they are quite alike, depending on the point of view.
I created a system that learns in real time. It is more or less an architecture definition quite like reinforcement learning using the interval estimation algorithm. But instead of have basic steps we plan to use learning. The system then learns to learn.
I sometimes compare it to the teaching of a child to play soccer. You do not specify every move he/she has to make. I would put the kiddo in front of a goal, and tell him that he should kick the ball in the goal, maybe showing it once, but not more than once. He starts kicking and after a while I look and give some hints about (for example) that he should put his foot a bit different. Then he starts to train again. So effectively I am teaching the kid how he can learn for himself how to solve the problem.
The MORPH proposal is accepted we can really start to take off. At the moment Axel Brink are using Evolutionary Computation for the pre-processing of images from color to black-and-white. The user only clicks on what he/she likes and the algorithm is being trained on the fly. After a while the user sees that it is processed correctly and this is used for the next time.
Nowadays I am focusing on RoboCup@Home again.